Parallel binocular stereo-vision-based GPU accelerated pedestrian detection and distance computation

Abstract

Pedestrian detection has become a very hot research field in computer vision, because it is widely used in many practical applications. However, the real-time requirement of these applications is a great challenge for pedestrian detection. To address this problem, this paper accelerates the pedestrian detection in parallel using NVIDIA’s Graphics Processing Units (GPUs). In addition, we developed a distance estimation system based on the results of the pedestrian detection, which aims to obtain the distance between the pedestrians and the camera. The whole system including pedestrian detection and distance estimation is for embedded applications. The method of pedestrian detection is to combine the Histogram of Oriented Gradients (HOG) feature with the cascade classifier, and the distance estimation system is built by utilizing a parallel binocular vision system. The performance of the parallel implementation of the whole system is tested on two kinds of different GPUs, an embedded board Jetson TK1 and a Tesla K80 GPU specialized for science computation. The speed of the whole system on Jetson TK1 over 640 × 480 images is about 16 fps, which basically reaches the real-time requirement, and the speed on Tesla K80 over 640 × 480 images is much higher, about 86 fps.